ITCNN: Incremental Learning Network Based on ITDA and Tree Hierarchical CNN

In class incremental learning, sensitive changes in network parameter tuning pose a significant problem. To address this challenge, we propose a new model called ITCNN, which combines a tree-like hierarchical network structure with incremental tensor discriminant analysis (ITDA). Unlike previous deep convolutional neural networks (DCNNs) that rely on data organization, ITCNN is feature-driven, using a tree hierarchy that can learn and grow with new data while retaining the classification functions of previous categories. The use of ITDA obviates the need to represent high-dimensional training samples in vector form, and the incremental computation has less space and time complexity. Compared with other incremental learning methods, our proposed network achieves competitive accuracy on multiple datasets. Our results suggest that ITCNN is an effective model for addressing the challenges of class incremental learning, providing increased accuracy while retaining efficiency and scalability.